lkml.org 
[lkml]   [2013]   [Mar]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] bounce:fix bug, avoid to flush dcache on slab page from jbd2.
On Fri, Mar 15, 2013 at 11:01:05AM +0100, Jan Kara wrote:
> On Thu 14-03-13 15:42:43, Darrick J. Wong wrote:
> > On Wed, Mar 13, 2013 at 10:02:16PM +0100, Jan Kara wrote:
> > > On Wed 13-03-13 12:44:29, Darrick J. Wong wrote:
> > > > On Wed, Mar 13, 2013 at 09:50:21AM +0100, Jan Kara wrote:
> > > > > On Tue 12-03-13 18:10:20, Darrick J. Wong wrote:
> > > > > > On Tue, Mar 12, 2013 at 03:32:21PM -0700, Andrew Morton wrote:
> > > > > > > On Fri, 08 Mar 2013 20:37:36 +0800 Shuge <shugelinux@gmail.com> wrote:
> > > > > > >
> > > > > > > > The bounce accept slab pages from jbd2, and flush dcache on them.
> > > > > > > > When enabling VM_DEBUG, it will tigger VM_BUG_ON in page_mapping().
> > > > > > > > So, check PageSlab to avoid it in __blk_queue_bounce().
> > > > > > > >
> > > > > > > > Bug URL: http://lkml.org/lkml/2013/3/7/56
> > > > > > > >
> > > > > > > > ...
> > > > > > > >
> > > > > > > > --- a/mm/bounce.c
> > > > > > > > +++ b/mm/bounce.c
> > > > > > > > @@ -214,7 +214,8 @@ static void __blk_queue_bounce(struct request_queue
> > > > > > > > *q, struct bio **bio_orig,
> > > > > > > > if (rw == WRITE) {
> > > > > > > > char *vto, *vfrom;
> > > > > > > > - flush_dcache_page(from->bv_page);
> > > > > > > > + if (unlikely(!PageSlab(from->bv_page)))
> > > > > > > > + flush_dcache_page(from->bv_page);
> > > > > > > > vto = page_address(to->bv_page) + to->bv_offset;
> > > > > > > > vfrom = kmap(from->bv_page) + from->bv_offset;
> > > > > > > > memcpy(vto, vfrom, to->bv_len);
> > > > > > >
> > > > > > > I guess this is triggered by Catalin's f1a0c4aa0937975b ("arm64: Cache
> > > > > > > maintenance routines"), which added a page_mapping() call to arm64's
> > > > > > > arch/arm64/mm/flush.c:flush_dcache_page().
> > > > > > >
> > > > > > > What's happening is that jbd2 is using kmalloc() to allocate buffer_head
> > > > > > > data. That gets submitted down the BIO layer and __blk_queue_bounce()
> > > > > > > calls flush_dcache_page() which in the arm64 case calls page_mapping()
> > > > > > > and page_mapping() does VM_BUG_ON(PageSlab) and splat.
> > > > > > >
> > > > > > > The unusual thing about all of this is that the payload for some disk
> > > > > > > IO is coming from kmalloc, rather than being a user page. It's oddball
> > > > > > > but we've done this for ages and should continue to support it.
> > > > > > >
> > > > > > >
> > > > > > > Now, the page from kmalloc() cannot be in highmem, so why did the
> > > > > > > bounce code decide to bounce it?
> > > > > > >
> > > > > > > __blk_queue_bounce() does
> > > > > > >
> > > > > > > /*
> > > > > > > * is destination page below bounce pfn?
> > > > > > > */
> > > > > > > if (page_to_pfn(page) <= queue_bounce_pfn(q) && !force)
> > > > > > > continue;
> > > > > > >
> > > > > > > and `force' comes from must_snapshot_stable_pages(). But
> > > > > > > must_snapshot_stable_pages() must have returned false, because if it
> > > > > > > had returned true then it would have been must_snapshot_stable_pages()
> > > > > > > which went BUG, because must_snapshot_stable_pages() calls page_mapping().
> > > > > > >
> > > > > > > So my tentative diagosis is that arm64 is fishy. A page which was
> > > > > > > allocated via jbd2_alloc(GFP_NOFS)->kmem_cache_alloc() ended up being
> > > > > > > above arm64's queue_bounce_pfn(). Can you please do a bit of
> > > > > > > investigation to work out if this is what is happening? Find out why
> > > > > > > __blk_queue_bounce() decided to bounce a page which shouldn't have been
> > > > > > > bounced?
> > > > > >
> > > > > > That sure is strange. I didn't see any obvious reasons why we'd end up with a
> > > > > > kmalloc above queue_bounce_pfn(). But then I don't have any arm64s either.
> > > > > >
> > > > > > > This is all terribly fragile :( afaict if someone sets
> > > > > > > bdi_cap_stable_pages_required() against that jbd2 queue, we're going to
> > > > > > > hit that BUG_ON() again, via must_snapshot_stable_pages()'s
> > > > > > > page_mapping() call. (Darrick, this means you ;))
> > > > > >
> > > > > > Wheeee. You're right, we shouldn't be calling page_mapping on slab pages.
> > > > > > We can keep walking the bio segments to find a non-slab page that can tell us
> > > > > > MS_SNAP_STABLE is set, since we probably won't need the bounce buffer anyway.
> > > > > >
> > > > > > How does something like this look? (+ the patch above)
> > > > > Umm, this won't quite work. We can have a bio which has just PageSlab
> > > > > page attached and so you won't be able to get to the superblock. Heh, isn't
> > > > > the whole page_mapping() thing in must_snapshot_stable_pages() wrong? When we
> > > > > do direct IO, these pages come directly from userspace and hell knows where
> > > > > they come from. Definitely their page_mapping() doesn't give us anything
> > > > > useful... Sorry for not realizing this earlier when reviewing the patch.
> > > > >
> > > > > ... remembering why we need to get to sb and why ext3 needs this ... So
> > > > > maybe a better solution would be to have a bio flag meaning that pages need
> > > > > bouncing? And we would set it from filesystems that need it - in case of
> > > > > ext3 only writeback of data from kjournald actually needs to bounce the
> > > > > pages. Thoughts?
> > > >
> > > > What about dirty pages that don't result in journal transactions? I think
> > > > ext3_sync_file() eventually calls ext3_ordered_writepage, which then calls
> > > > __block_write_full_page, which in turn calls submit_bh().
> > > So here we have two options:
> > > Either we let ext3 wait the same way as other filesystems when stable pages
> > > are required. Then only data IO from kjournald needs to be bounced (all
> > > other IO is properly protected by PageWriteback bit).
> > >
> > > Or we won't let ext3 wait (as it is now), keep the superblock flag that fs
> > > needs bouncing, and set the bio flag in __block_write_full_page() and
> > > kjournald based on the sb flag.
> > >
> > > I think the first option is slightly better but I don't feel strongly
> > > about that.
> >
> > I like that first option -- it contains the kludgery to jbd instead of
> > spreading it around. Here's a patch that passes a quick smoke test on ext[34],
> > xfs, and vfat. What do you think of this one? Should I create a
> > submit_snapshot_bh() instead of letting callers stuff in arbitrary dangerous
> > BH_ flags?
> Thanks for writing the patch. I think _submit_bh() is OK as you did it. I
> have just two comments below.
>
> > ---
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > Subject: [PATCH] mm: Make snapshotting pages for stable writes a per-bio operation
> >
> > Walking a bio's page mappings has proved problematic, so create a new bio flag
> > to indicate that a bio's data needs to be snapshotted in order to guarantee
> > stable pages during writeback. Next, for the one user (ext3/jbd) of
> > snapshotting, hook all the places where writes can be initiated without
> > PG_writeback set, and set BIO_SNAP_STABLE there. Finally, the MS_SNAP_STABLE
> > mount flag (only used by ext3) is now superfluous, so get rid of it.
> >
> > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > ---
> > fs/buffer.c | 9 ++++++++-
> > fs/ext3/super.c | 1 -
> > fs/jbd/commit.c | 4 ++--
> > include/linux/blk_types.h | 3 ++-
> > include/linux/buffer_head.h | 1 +
> > include/uapi/linux/fs.h | 1 -
> > mm/bounce.c | 21 +--------------------
> > mm/page-writeback.c | 4 ----
> > 8 files changed, 14 insertions(+), 30 deletions(-)
> >
> ...
> > diff --git a/fs/jbd/commit.c b/fs/jbd/commit.c
> > index 86b39b1..b91b688 100644
> > --- a/fs/jbd/commit.c
> > +++ b/fs/jbd/commit.c
> > @@ -163,7 +163,7 @@ static void journal_do_submit_data(struct buffer_head **wbuf, int bufs,
> > for (i = 0; i < bufs; i++) {
> > wbuf[i]->b_end_io = end_buffer_write_sync;
> > /* We use-up our safety reference in submit_bh() */
> > - submit_bh(write_op, wbuf[i]);
> > + _submit_bh(write_op, wbuf[i], 1 << BIO_SNAP_STABLE);
> Please add a comment here why we need BIO_SNAP_STABLE. Something like:
> /*
> * Here we write back pagecache data that may be mmaped. Since we cannot
> * afford to clean the page and set PageWriteback here due to lock ordering
> * (page lock ranks above transaction start), the data can change while IO is
> * in flight. Tell the block layer it should bounce the bio pages if stable
> * data during write is required.
> */
>
> > }
> > }
> >
> > @@ -667,7 +667,7 @@ start_journal_io:
> > clear_buffer_dirty(bh);
> > set_buffer_uptodate(bh);
> > bh->b_end_io = journal_end_buffer_io_sync;
> > - submit_bh(write_op, bh);
> > + _submit_bh(write_op, bh, 1 << BIO_SNAP_STABLE);
> And this isn't needed. Here we write out only metadata and JBD already
> handles copying those / waiting for IO in flight for metadata.

I think it only copies the page if either the buffer is also a part of the
current transaction (or someone called do_get_undo_access()). Unfortunately,
if we're in data=journal mode, dirty data pages get pushed through jbd as if
they were fs metadata, but in the meantime other processes can still write to
those pages. So I guess we need the journal to freeze those pages as soon as
they come in.

(Or we could retain that little piece, but I suppose it's a larger hammer than
necessary.)

--D
>
> The rest of the patch looks OK and I like it much more than the previous
> version :)
>
> Honza
> > }
> > cond_resched();
> >
> > diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
> > index cdf1119..22990cf 100644
> > --- a/include/linux/blk_types.h
> > +++ b/include/linux/blk_types.h
> > @@ -111,12 +111,13 @@ struct bio {
> > #define BIO_FS_INTEGRITY 9 /* fs owns integrity data, not block layer */
> > #define BIO_QUIET 10 /* Make BIO Quiet */
> > #define BIO_MAPPED_INTEGRITY 11/* integrity metadata has been remapped */
> > +#define BIO_SNAP_STABLE 12 /* bio data must be snapshotted during write */
> >
> > /*
> > * Flags starting here get preserved by bio_reset() - this includes
> > * BIO_POOL_IDX()
> > */
> > -#define BIO_RESET_BITS 12
> > +#define BIO_RESET_BITS 13
> >
> > #define bio_flagged(bio, flag) ((bio)->bi_flags & (1 << (flag)))
> >
> > diff --git a/include/linux/buffer_head.h b/include/linux/buffer_head.h
> > index 5afc4f9..714d5d9 100644
> > --- a/include/linux/buffer_head.h
> > +++ b/include/linux/buffer_head.h
> > @@ -181,6 +181,7 @@ void ll_rw_block(int, int, struct buffer_head * bh[]);
> > int sync_dirty_buffer(struct buffer_head *bh);
> > int __sync_dirty_buffer(struct buffer_head *bh, int rw);
> > void write_dirty_buffer(struct buffer_head *bh, int rw);
> > +int _submit_bh(int, struct buffer_head *, unsigned long);
> > int submit_bh(int, struct buffer_head *);
> > void write_boundary_block(struct block_device *bdev,
> > sector_t bblock, unsigned blocksize);
> > diff --git a/include/uapi/linux/fs.h b/include/uapi/linux/fs.h
> > index c7fc1e6..a4ed56c 100644
> > --- a/include/uapi/linux/fs.h
> > +++ b/include/uapi/linux/fs.h
> > @@ -88,7 +88,6 @@ struct inodes_stat_t {
> > #define MS_STRICTATIME (1<<24) /* Always perform atime updates */
> >
> > /* These sb flags are internal to the kernel */
> > -#define MS_SNAP_STABLE (1<<27) /* Snapshot pages during writeback, if needed */
> > #define MS_NOSEC (1<<28)
> > #define MS_BORN (1<<29)
> > #define MS_ACTIVE (1<<30)
> > diff --git a/mm/bounce.c b/mm/bounce.c
> > index 5f89017..a5c2ec3 100644
> > --- a/mm/bounce.c
> > +++ b/mm/bounce.c
> > @@ -181,32 +181,13 @@ static void bounce_end_io_read_isa(struct bio *bio, int err)
> > #ifdef CONFIG_NEED_BOUNCE_POOL
> > static int must_snapshot_stable_pages(struct request_queue *q, struct bio *bio)
> > {
> > - struct page *page;
> > - struct backing_dev_info *bdi;
> > - struct address_space *mapping;
> > - struct bio_vec *from;
> > - int i;
> > -
> > if (bio_data_dir(bio) != WRITE)
> > return 0;
> >
> > if (!bdi_cap_stable_pages_required(&q->backing_dev_info))
> > return 0;
> >
> > - /*
> > - * Based on the first page that has a valid mapping, decide whether or
> > - * not we have to employ bounce buffering to guarantee stable pages.
> > - */
> > - bio_for_each_segment(from, bio, i) {
> > - page = from->bv_page;
> > - mapping = page_mapping(page);
> > - if (!mapping)
> > - continue;
> > - bdi = mapping->backing_dev_info;
> > - return mapping->host->i_sb->s_flags & MS_SNAP_STABLE;
> > - }
> > -
> > - return 0;
> > + return test_bit(BIO_SNAP_STABLE, &bio->bi_flags);
> > }
> > #else
> > static int must_snapshot_stable_pages(struct request_queue *q, struct bio *bio)
> > diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> > index efe6814..4514ad7 100644
> > --- a/mm/page-writeback.c
> > +++ b/mm/page-writeback.c
> > @@ -2311,10 +2311,6 @@ void wait_for_stable_page(struct page *page)
> >
> > if (!bdi_cap_stable_pages_required(bdi))
> > return;
> > -#ifdef CONFIG_NEED_BOUNCE_POOL
> > - if (mapping->host->i_sb->s_flags & MS_SNAP_STABLE)
> > - return;
> > -#endif /* CONFIG_NEED_BOUNCE_POOL */
> >
> > wait_on_page_writeback(page);
> > }
> --
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html


\
 
 \ /
  Last update: 2013-03-15 19:21    [W:0.091 / U:0.940 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site